Chronos 2
autogluon/chronos-2
published Oct 2025 · updated Jun 2026
Chronos 2 is a time-series-forecasting model that provides zero-shot univariate, multivariate, and covariate-informed forecasting with state-of-the-art accuracy.
specs
| Task | Time Series Forecasting |
| Architecture | Encoder-only transformer with group attention (inspired by T5) |
| Parameters | 120M |
| License | Apache-2.0 |
| Max Context Length | 8192 |
| Max Prediction Length | 1024 |
about this model
Chronos-2 is a 120M-parameter encoder-only time series foundation model for zero-shot forecasting that supports univariate, multivariate, and covariate-informed tasks within a single architecture. Inspired by the T5 encoder, it uses a group attention mechanism to enable efficient in-context learning across related series and covariates, producing multi-step-ahead quantile forecasts.
Capabilities
| Capability | Chronos-2 |
|---|---|
| Univariate forecasting | ✅ |
| Cross-learning across items | ✅ |
| Multivariate forecasting | ✅ |
| Past-only covariates (real/categorical) | ✅ |
| Known future covariates (real/categorical) | ✅ |
| Max. context length | 8192 |
| Max. prediction length | 1024 |
Performance
Chronos-2 achieves state-of-the-art zero-shot accuracy among public models on fev-bench, GIFT-Eval, and Chronos Benchmark II. In head-to-head comparisons it achieves a win rate of over 90% against Chronos-Bolt. Its largest performance improvements over baselines are observed on tasks that include exogenous features (covariates). The model delivers over 300 time series forecasts per second on a single A10G GPU and supports both GPU and CPU inference.
Additional Details
Trained on a combination of real-world and large-scale synthetic datasets. Case studies in the technical report demonstrate practical advantages in energy and retail domains. Licensed under Apache-2.0. For more information, see the GitHub repository and technical report.
best for
- ·Zero-shot univariate forecasting without task-specific training
- ·Multivariate forecasting with cross-series learning
- ·Covariate-informed forecasting (e.g., energy demand with weather data)
- ·High-throughput batch forecasting (over 300 forecasts/sec on A10G)
FAQ
Chronos-2 excels at zero-shot forecasting for univariate, multivariate, and covariate-informed tasks, achieving state-of-the-art accuracy on fev-bench, GIFT-Eval, and Chronos Benchmark II.
Chronos-2 supports multivariate and covariate-informed forecasting, while Chronos-Bolt does not. Chronos-2 also has a larger context length (8192 vs 2048) and achieves a win rate over 90% against Chronos-Bolt.
Chronos-2 is released under the Apache-2.0 license, which allows free use, modification, and distribution.
Chronos-2 accepts time series data as pandas DataFrames with columns for target, timestamp, and optional covariates. In production, it can also accept JSON payloads with target values and parameters.
Chronos-2 is available as a hosted API on gigarouter. Use the OpenAI-compatible endpoint with your API key to send requests in JSON format and receive forecasts.
We're benchmarking and onboarding Chronos 2 as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.